AI agent costs are emerging as a new financial challenge for health systems racing to automate clinical and administrative workflows. Rush University System for Health has spent more than 18 months developing AI agents, but its CIO says the industry still lacks a reliable way to model what those agents will cost before they are activated. Unlike traditional software licenses with relatively predictable subscription fees, agent expenses can rise according to how often systems run, how much information they process and which artificial intelligence models they use.
The result is a potentially significant gap between an AI project’s expected savings and its actual operating cost.
AI Agent Costs Are Difficult to Predict
Health systems have traditionally budgeted technology around software licenses, infrastructure and staffing.
AI agents introduce a different financial model.
Many AI platforms charge according to token consumption. Tokens represent units of information processed or generated by an artificial intelligence model. A simple interaction may consume relatively few tokens, while an agent completing a complex multi-step workflow may repeatedly analyze large amounts of context, call external tools and generate additional responses.
Rush CIO Jeff Gautney warned that agents could eventually cost more than the human work they are intended to replace if organizations fail to establish financial discipline.
The problem becomes more difficult because healthcare organizations increasingly use AI capabilities from multiple technology vendors.
AI Agent Costs Lack Unified Visibility
Nearly every major enterprise platform is introducing tools for building AI agents.
A health system could eventually operate agents through its EHR, customer relationship management platform, cloud provider, revenue-cycle systems and other applications.
Each vendor may report token usage differently and expose that information through its own administrative dashboard.
Rush says there is currently no common mechanism for combining those costs into a single financial model.
The FinOps Foundation identifies the same broader problem. Its 2026 research found that AI cost management has become the top skill organizations believe their FinOps teams need to develop, while 98% of surveyed organizations now manage some AI spend.
Healthcare Needs AI FinOps Discipline
One potential solution is applying the financial operations practices already used for cloud computing.
Cloud FinOps teams monitor consumption, allocate spending to business units, forecast demand and identify unnecessary usage.
AI requires a similar model.
The FinOps Foundation recommends tracking token usage and attributing consumption to individual applications or business use cases rather than treating AI spending as one large enterprise technology expense.
Health systems could therefore track the cost per prescription refill completed, patient inquiry answered or authorization processed.
That makes it possible to compare AI expense with measurable operational value.
AI Agent Costs Require Model Selection
One of the easiest ways to increase AI spending is using an unnecessarily powerful model.
Gautney said employees often want to select the most advanced frontier model because they assume it must produce the best result. But those models may be considerably more expensive than simpler alternatives.
A routine workflow such as finding clinic hours may require little reasoning capability.
Using an advanced model for that task could increase expenses without materially improving the result.
Some processes may not need generative AI at all. Traditional search, rules-based automation or smaller models could accomplish the same objective more efficiently.
The FinOps Foundation recommends this type of model right-sizing, describing it as one of the highest-leverage interventions for organizations trying to control AI spending.
Rush Uses Training as Financial Control
Because cost-management technology remains immature, Rush is concentrating on the people authorized to build agents.
Its AI education program uses several training levels.
A foundational course is required for IT staff. A second level is aimed at citizen developers who receive access to governed datasets and agent-building tools, while a more advanced level is designed for employees working with AI systems regularly.
The objective is not simply teaching employees how to build agents.
Teams also need to understand when agents are appropriate, which models should be selected and what financial consequences their design choices create.
AI Agent Costs Extend Beyond Tokens
Token consumption is only one part of the total cost.
An AI agent is functioning software that must continue operating after its original developer moves to another job or leaves the organization.
Underlying systems can change. APIs can be modified. Models may be updated. Workflows may evolve.
An agent created by someone in supply chain, finance or clinical operations could continue running long after the employee who built it has departed.
Rush argues that this creates a new ownership problem: when the agent eventually breaks, someone must know who is responsible for maintaining it.
Health systems may eventually operate hundreds or even thousands of agents, making lifecycle management a major IT responsibility.
Agents Need Ownership From Launch
Every production AI agent should therefore have an identifiable owner.
Organizations should document its purpose, model, expected usage, cost center, dependencies, escalation path and business sponsor.
They should also determine how the agent will be retired when it is no longer useful.
This is similar to application portfolio management, but AI agents may proliferate much faster because modern platforms increasingly allow nontechnical employees to create them.
Without governance, organizations could accumulate large numbers of lightly monitored agents consuming resources in the background.
The FinOps Foundation describes this emerging discipline as token economics, or Tokenomics—connecting AI consumption with business outcomes so organizations can determine whether spending is actually creating value.
AI Agent Costs Must Connect to ROI
The central question for health systems should not be whether an agent is technically impressive.
It should be whether the agent generates enough measurable value to justify its full cost.
An agent processing prescription refills, for example, could be evaluated on cost per completed request, employee time saved, turnaround time, patient satisfaction and escalation rates.
A patient-service agent could be measured against call-center workload and resolution rates.
These unit economics make AI investments easier for finance and clinical leaders to evaluate.
The FinOps Foundation warns that advertised per-token pricing alone does not represent true AI cost. Context growth, model selection and workflow structure can substantially change spending, making use-case economics more important than simply counting tokens.
Healthcare AI Needs Financial Governance
AI agents may ultimately provide substantial productivity gains across scheduling, revenue cycle, patient communication, documentation and clinical operations.
But scaling them without financial visibility could create a new category of uncontrolled technology spending.
Health systems should therefore establish cost attribution, model-selection standards, usage monitoring, ownership rules and retirement processes before agent adoption accelerates.
The concern surrounding AI agent costs does not mean hospitals should slow useful innovation. It means agentic AI needs the same financial discipline healthcare organizations eventually developed around cloud computing.
As AI agents become easier for employees to create, the organizations that succeed will likely be those capable of answering a simple question for every deployment: what does this agent cost, who owns it and what measurable value does it produce?
For more healthcare industry updates, insights and news, visit DistilINFO. Click here to subscribe to stay informed
